Selected Topics in Biostatistics Seminar Series Association and Causation Sponsored by: Center For Clinical Investigation and Cleveland CTSC Vinay K. Cheruvu, MSc., MS Biostatistician, CTSC BERD cheruvu@case.edu April 28, 2010
OUTLINE Introduction to Association & Causation Conceptual framework Confounding & Interaction Methods to study Association & Causation Remarks Take home message Questions 2
SCIENTIFIC QUESTION Exposure Outcome(Disease) Family of Other Third Variable(s) Questions: 1. Association or Causation??? 2. How is the family of Third variable(s) related to exposure or(and) Outcome? 3
ASSOCIATION Association (relationship between exposure and outcome) implies the two are not independent but related in some way or the other. A very basic measure of association is correlation (strength or degree of association). No need to ascertain temporality. Association does not necessarily imply causation. It is just an identifiable relationship. 4
CAUSATION A true mechanism that leads from exposure to outcome i.e., change in exposure measured directly causes a change in the outcome. It is an attempt to establish scientifically a valid causal link(connection) between exposure & outcome. If we determine association, then a logical question is whether the observed association reflects a causal relationship. Need to ascertain temporality (cause must precede the effect). 5
CAUTION COMES WITH CAUSATION A statistical ti ti relationship, however strong and however suggestive, can never establish causal connection: our ideas of causation must come from outside statistics, ultimately t l from some theory or other. ** Without considerable additional research, replication of studies, and experimentation, results claiming a causal relationship could be misleading. ** M. G. Kendall and A. Stuart, The Advanced Theory of Statistics, Charles Griffin Publishers, New York, 1961, vol. 2, chap. 26, p. 279. 6
CRITERIA FOR CAUSATION Criteria originally proposed by Austin Bradford Hill (1897-1991) Temporality (exposure always precedes the outcome). Strength of association (degree of relative risk). Dose-Response (increase in exposure is associated with increase in risk). Consistency (studies replicated in different settings show similar(consistent) results). 7
CRITERIA FOR CAUSATION contd. Criteria originally proposed by Austin Bradford Hill (1897-1991) Biological Plausibility (sound theoretical basis and knowledge about the process). Alternative Explanations (judge if association determined is causal). Experiment & Coherence (results compatible with current ttheory and dknowledge). 8
CONCEPTUAL FRAMEWORK Different types of study designs - Randomized trials(experiments) - Observational studies Feasibility - Is it ethical to conduct the study? - Nature of exposure? 9
CONFOUNDING Exposure Outcome(Disease) Family of other Third variable(s) The third variable(factor) is a confounder if - the factor is a known risk for outcome(disease). - the factor is associated with exposure but is not a result of exposure. 10
Example CONFOUNDING Coffee Drinking Risk of pancreatic cancer Age, Smoking Considering the definition of confounding - We know both Age, Smoking is a risk for cancer. - We know Age, Smoking is associated with Coffee Drinking but not a result of it. 11
INTERACTION (EFFECT MODIFICATION) Exposure Outcome(Disease) Family of other Third variable(s) The third variable(factor) is a effect modifier if - the relationship between the exposure and outcome is different among different levels(subgroups) of the factor. - In other words, the effect of exposure on outcome depends on the level of the factor. 12
INTERACTION (EFFECT MODIFICATION) Coffee Drinking Risk of pancreatic cancer Gender Considering the definition of effect modifier - We know gender is related to risk of pancreatic cancer. - The effect of coffee drinking on risk of pancreatic cancer could be different by gender. 13
HOW TO ASSESS CONFOUNDING AND INTERACTION Consider a hypothetical study with the following information: A total of 200 subjects in the study of which 100 have pancreatic cancer(cases) and 100 don t have pancreatic cancer(controls); 48 subjects drink coffee(exposed) and 152 subjects don t drink coffee(unexposed); 130 subjects are smokers and 70 subjects are non-smokers; 90 subjects are above age 50(years) and 110 subjects are less than 50 years old; 95 males aesand 105 females. aes Cases Controls Exposed 30 18 Cases Controls Smoking 50 80 Unexposed 70 82 No Smoking 50 20 Cases Controls Age50+ 52 38 Age < 50 48 62 Cases Controls Males 51 44 Females 49 56 14
HOW TO ASSESS CONFOUNDING AND INTERACTION Complete Confounding Cases Controls Coffee Drinking 30 18 No Coffee Drinking 70 82 Crude RR = 1.36 Smoking No Smoking Cases Controls Cases Controls Coffee Drinking 5 8 No Coffee Drinking 45 72 Stratified RR = 1.0 Coffee Drinking 25 10 No Coffee Drinking 25 10 Stratified RR = 1.0 Source: Leon Gordis, Epidemiology, 2 nd edition. 15
HOW TO ASSESS CONFOUNDING AND INTERACTION No Confounding Cases Controls Coffee Drinking 30 18 No Coffee Drinking 70 82 Crude RR = 1.36 Age 50+ Age < 50 Cases Controls Cases Controls Coffee Drinking 22 12 No Coffee Drinking 30 26 Stratified RR = 1.21 Coffee Drinking 8 6 No Coffee Drinking 40 56 Stratified RR = 1.37 16
HOW TO ASSESS CONFOUNDING AND INTERACTION Presence of Interaction (Effect Modification) Cases Controls Coffee Drinking 30 18 No Coffee Drinking 70 82 Crude RR = 1.36 Males Females Cases Controls Cases Controls Coffee Drinking 16 2 No Coffee Drinking 35 42 Stratified RR = 1.96 Coffee Drinking 14 16 No Coffee Drinking 35 40 Stratified RR = 1.00 17
HOW TO HANDLE CONFOUNDING AND INTERACTION Confounding - Regression approach(adjustment) - Matching - Propensity scores - Design stage(covariate adaptive randomization) Interaction - Regression approach - Stratified analysis 18
METHODS TO STUDY ASSOCIATION AND CAUSATION Association - Regression approach Causation - Regression approach Structural Equation Modeling Marginal Structural Models Potential Outcomes G-estimation 19
REMARKS Association does not necessarily imply causation. It is just an identifiable relationship. Nevertheless, association almost always leads to pursue if a causal relationship exists. In order to establish a causal link, the exposure should always precede the outcome. Several methods exist to study association and causation(depending on how we define the measure of association or causation). Confounding and Interaction are two different concepts though the approach to assess its presence is simultaneous. Several methods exist to handle confounding & interaction. 20
TAKE HOME MESSAGE Source: www.xkcd.com, Thanks to Brian Schmotzer 21
QUESTIONS? 22